Every agent is a typed function over a pinned Anthropic model with a Zod-validated output schema. Read-only agents run automatically; HITL-gated agents draft a proposal that no integration touches until a licensed reviewer approves — that click writes the Submission audit row. Fire a scenario beat below and watch the agents react in real time.
Drafted by playbooks / AI · NOT sent until you approve. Approve fires the integration mock + writes a Submission audit row.
Implemented and audit-logged via runAgent()
Summarizes the inpatient record into 3 barrier candidates, predicted disposition, and risk flags.
Drafts a complete pharmacy PA packet — drug, justification, attempted alternatives, urgency, citations.
Tool-use chat that answers questions and proposes HITL-gated actions into the Pending Actions queue.
Declared in the AgentName union; prompts + tool schemas pending. Listed for transparency, not as shipped capability.
Outbound voice calls to SNFs / DME vendors to confirm capacity, bed-hold windows, and intake clinicians.
Pre-warms a backup resource (SNF, HH agency, DME) when the primary match is at-risk.
Re-renders clinical updates into 6th-grade reading level in the patient's preferred language.
Captures pre-admit functional + caregiver context from the patient/family via portal or call.
Plain-language explanation of why a clinical or operational alert fired, with the citation chain.
Per CLAUDE.md: every agent call goes through runAgent in src/lib/ai/audit.ts — no exceptions. The Cockpit Copilot uses the Anthropic tool-use loop directly, but still writes one AgentInference row per chat turn.